LFM2.5 2.6b vs MiniCPM5 2b

I tried both these models on a few small agentic tasks with tools (i.e. "What's the weather like today?", etc.). They're both pretty solid at using web search to find answers despite being small models.

TLDR:

LFM2.5 2.6b is the clear winner. Somehow it's faster and uses less ram than MiniCPM despite having more parameters. It also seems better aligned for english conversation.

MiniCPM5

• On an M1 air: pp 162 t/s - tg 16 t/s
• Uses about 3.8gb of ram at 32k context (with draft model)
• Often responds in chinese despite my prompting in english.
• It's very smart when it does respond in english and may be stronger at agentic work.
• It uses more memory than LFM2.5 despite supposedly having less parameters.
• There's a corresponding dspark model available.

​

llama-server --model MiniCPM5-2B-Q8_0.gguf -md MiniCPM5-2B-DSpark-Q8_0.gguf --load-mode none --spec-type draft-dspark --spec-draft-n-max 2 -ngl all -ngld all -fa on -np 1 -t 4 -c 32000 --reasoning on -fit off --temp 1.0 --top-p 0.95 --cache-type-k q5_1 --cache-type-v q5_1 --spec-draft-type-k q8_0 --spec-draft-type-v q8_0 

LFM2.5

• On an M1 air: pp 200 t/s - tg 22 t/s
• Uses about 2.5gb of ram at 32k context
• Works well for simple one shot agentic work but tends to start hallucinating quickly as the conversation gets longer.

​

llama-server -m LFM2.5-2.6B-QAD-Q4_0.gguf -ngl all -fa on --load-mode none --temp 0.1 --top-k 50 --top-p 0.9 -c 32000 --threads 4 --reasoning on -fit off --reasoning-preserve